OpenAI Cuts GPT-5.6 Sol Prices, Gemini 3.7 Flash Breaks Growth Record, NVIDIA Vera Rubin Enters Production

Key Takeaways

1. OpenAI will cut GPT-5.6 Sol API and credit prices by over 20% within three months, aiming for market-lowest pricing and top capability ceiling.
2. Gemini 3.7 Flash set a first-week growth record for Gemini models and is now in Search and the Gemini app.
3. NVIDIA Vera Rubin entered full production, with initial units arriving at a Microsoft datacenter.
4. Runway Ruby launched for Max and Enterprise, adding high-end 16-bit EXR, 10/12-bit ProRes/HEVC, BT.2020, and PQ/HLG support.
5. Agent workflows are maturing quickly — from turning real AI failures into reusable evals to autonomous bug-bounty hunting and low-risk task automation.

1. Models, Pricing, and Product Updates

1. OpenAI announced that GPT-5.6 Sol API and credit pricing will drop by more than 20% over the next three months. Greg Brockman cited OpenAI’s goal to offer the market’s lowest price and highest capability ceiling, and Sam Altman flagged the update as well. — via 1 2 3

2. Demis Hassabis said Gemini 3.7 Flash broke previous Gemini growth records in its first week and is now applied in Search and the Gemini app. The milestone points to fast adoption of the new Flash tier. — via 1

3. OpenAI’s weekly ChatGPT update brings faster recent-photo attachments on iOS, improved time understanding, quicker long-conversation loading, and optimized network error messages. These are small but practical usability improvements for everyday ChatGPT use. — via 1

4. Runway made Ruby available to Max and Enterprise users, converting any video to 16-bit EXR sequences, 10/12-bit ProRes/HEVC, BT.2020 color space, and PQ/HLG. The release targets high-end production and post-production workflows. — via 1

2. Infrastructure, Research, and Expansion

1. NVIDIA Vera Rubin entered full production, with the first production units arriving at a Microsoft datacenter. NVIDIA thanked Microsoft and Azure teams, signaling a concrete delivery step for next-generation infrastructure. — via 1

2. NVIDIA also argued that AI agent safety should come from a harness steering the agent plus infrastructure limiting its capabilities, viewed across the whole agent stack. This frames safety as a system design issue, not just a model-level property. — via 1

3. Demis Hassabis called games a key AI research testbed for 15 years, from Atari to StarCraft II, and says work with FenrisCreations targets continual learning, deep memory, long-term planning, and multi-agent dynamics. The long-term goal is using AI to create new game experiences and apply those advances to real-world and scientific discovery. — via 1

4. Runway is entering Latin America, with meetings in Chile involving enterprises, clients, and government, plus local community events. Runway says LATAM is already a meaningful user base and continues its global push after Japan, France, and the UK. — via 1

3. Agents, Workflows, and Adoption Signals

1. Ethan Mollick says Codex and Claude Code can now complete low-risk, time-consuming tasks like filling in forms from email without user intervention. This shows coding agents are becoming practical for everyday productivity, not just complex software work. — via 1

2. Deedy shared the example of a young person who didn’t attend college using 50+ AI agents, cybersecurity tools, and CVE indexes to attempt 24/7 to break production sites and earn bug bounties. The story suggests AI-augmented security testing can be a viable solo career path. — via 1

3. Hamel Husain highlighted a new show segment that live-critiques AI evals and shows how to use a free Error Discovery skill in Claude Code or Codex to turn real AI failures into reusable evals. It is a practical resource for teams building LLM evaluation pipelines. — via 1

4. swyx argues that if you haven’t set Codex, Claude, Gemini, or Devin to automatically research SEO/AEO improvements each week, you’re missing free and unexplored alpha. The takeaway is to let agents run recurring optimization research rather than doing it manually. — via 1

5. Ben Tossell suggested using Grok Bot for long outputs by having it publish to an external site and return a link, rather than filling the chat window. He also criticized Claude for not supporting local session creation from mobile, arguing this pushes users to Codex. — via 1 2

4. Policy, Society, and AI Adoption

1. Yann LeCun amplified criticism that the Trump administration promised not to cut Medicaid but is cutting nearly $1 trillion, leaving over 5 million people without health insurance. The post highlights how AI leaders are engaging with U.S. policy impacts outside technology. — via 1

2. Ethan Mollick predicts a contradictory era in which polls show people broadly dislike AI, yet many secretly continue using it. He notes AI companies hold a public-opinion disadvantage while users develop strong attachments to favored models. — via 1